> Fine grain locking may be 10x slower than with no > locking, but as I recall from our mutual reading and > discussions w.r.t. B-link trees, updaters only block > each other (not readers)
The way I had to code it, readers still need to do a synchronized check if they are blocked by a writer. The readers pay the 10X cost just for checking. That is not counting being blocked. It there is a way for the readers to not do a check it would go faster. Any suggestions would be very interesting. > and only on competing for > individual node locks, unlike some tree That is basically why the cost higher. The check is per node so an read requires about 4 checks (best case (actaully 7 if we are using the cache)) vs a course grain lock would only require 1 check It is the read_lock_aquire() that is sucking up all the time . > implementations that require require whole path > locks to do updates (i.e. B-link trees are highly > localized w.r.t. rebalancing, as opposed to > pessimistic locking under T-Trees which tends to > lock the entire path from the root to the node being > updated). As such the overall number of locks is > still lower than for many other approaches, and That is why I am currently leaning toward coarse locking. It is the least amount of locking. > using the lateral traverse link pointer, readers > sail through, guaranteed that they will always > eventually find the correct leaf node (if it exists) yes it is in the code I implemented. > or properly fail the search. > > So I would suspect that until a very high % of > activity becomes both read and update activity > against a particular index in a particular table, > this 10x overhead will really represent a much > smaller overhead in terms of the overall time to possibly true. If a index lookup is only 1% of the time spend dealing with the query, then inflating it by 10 to get finegrain locking may be ok. Depends on the relative costs in the rest of the system. > process any particular statement. OTOH, what this > does buy is virtually unblocked multiple concurent > readers under multiple concurent writers, allowing a > mix of short and long running queries to experience > better perceived throughput; I think this will be yes. fine grain locking is 'fairer' to multiple threads. If a very long query starts and then a short read starts, the short read will not be greatly held up by the long query. But that price seems to be all index operation go 10X slower. >And when > developing software for the typical end-user, it is > alway good to be able to provide a perceived > increased responsiveness that is, within reason, > closer to what the end-user expects than was > previously being delivered. It would be possible to make it configurable. Maybe turn on and off fine grain checks. That would give a developer a choice. On the other hand BLinks are much more complicated to code. I still don't have a good delete function > > On another line of thought, straight BTrees may end > up being even slower than B-link under concurrent > access, I will let you know, I should have a BTree varient sometime next week. > using the same level of granularity (just a > thought...since you've already implmented B-link. > OTOH, maybe not). Have you done the research or > analysis yourself to determine this one way or > another? > > Responding to another point, perhaps some middle > ground can be found for granularity that lies One idea I had was that updates completely lock the table but all reads were unsynchronized. This gives you single check of course grain, but at least allows lots of reads to occure. > between locking single nodes and locking entire > high-level structures? Certainly, there are > opportunities, for example, to lazily rebalance, BTrees would not change depth in a rebalance but they would be able to release deleted memory. > concurrent vs thread-safe multiplexed) operation can > be switched, based on load. That is, with a single yes but complicated. Dont forget now you have added another synchronized check to what the current mode is ===== Karl Meissner ........................................................... : Meissner Software Development, LLC : : "Software development and project management services." : : http://meissner.v0.net/resume.htm : ........................................................... __________________________________________________ Do You Yahoo!? Yahoo! Games - play chess, backgammon, pool and more http://games.yahoo.com/ _______________________________________________ hsqldb-developers mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/hsqldb-developers
